14-01-2021 дата публикации
Номер: US20210012094A1
A two-stage person searching method combining face and appearance features, comprises: detecting a face of a person utilizing a face detector, and outputting a face representation vector based on a face comparing model; ranking person sets to be matched according to an Euclidean distance to acquire a face ranking result; selecting a plurality of samples as multi-matching targets at the next stage according to the ranking result; using the selected multi-matching targets of different persons at the next stage in the same data set as mutual negative samples, so as to compress the sample space matched at the next stage; and finally re-recognizing multi-target persons, and ranking the image sets to be matched according to an average distance or similarity with multiple targets to output a final result. 1. A two-stage person searching method combining face and appearance features , comprising:{'sub': 'C', 'acquiring an image I(x, y) containing a target person;'}{'sub': 'G', 'acquiring a panoramic image to be identified and person coordinate information in the panoramic image, and determining an image I(X, y) containing a candidate person;'}{'sub': C', 'G, 'calculating a score of face similarity between the target person and the candidate person according to the image I(x y) and the image I(x, y);'}{'sub': 'Q', 'ranking according to scores of face similarity, if the score is larger than or equal to a preset threshold, using corresponding image containing the candidate person as a target image I(x, y) for person re-identification; and if the score is less than the preset threshold, using the corresponding image containing the candidate person as the panoramic image for person re-identification;'}{'sub': G', 'C, 'filtering each panoramic image for person re-identification corresponding to the target person to obtain a processed candidate person image I′(x, y) if the image I(x, y) contains two or more target persons;'}{'sub': Q', 'G', 'Q', 'G, 'calculating an initial ...
Подробнее